Conclusion: the future of generative AI
Recap the six families and where things are heading.
This series went from a simple p(x) to systems that draw pictures from text. Every method differs in how it handles p(x): approximate it, compute it exactly, or skip it and just learn to sample.
Where things are heading: models across many modalities (text, image, audio, video), AI as a working assistant, and ethical questions such as copyright and misinformation.
All six families on one page
| Family | Strength | Weakness |
|---|---|---|
| VAE | Easy to train, nice latent space | Blurry samples |
| GAN | Sharp images, fast sampling | Hard to train, mode collapse |
| Autoregressive | Exact likelihood, great for text | Slow generation |
| Normalizing flow | Exact likelihood + invertible | Constrained network |
| Energy-based | Flexible model form | Slow sampling |
| Diffusion | Top quality, stable training | Many sampling steps |
Next steps
- Run every lesson's code, then swap in your own data
- Try Hugging Face diffusers to run the real Stable Diffusion
- Use generative AI responsibly: credit sources and respect copyright
- #GenerativeAI
- #DeepLearning
- #PythonສຳລັບຄົນLaos
- #AI
- #ບອກຕໍ່ຄວາມຮູ້